Changelog

0.1.4 · 2026-10-11

Agent management and a standalone, Node.js-free runtime.

• Agent management via MCP: Connect to /mcp using Streamable HTTP and existing scoped API keys. Query catalogs, manage drafts, preview Pipelines, publish, control flows and configure destinations, Workers and business settings. Existing validation, redaction, revision checks and activity records remain in effect.
• MCP settings: Enable MCP in System settings → Management API; it is off by default. The saved switch survives restarts. Read-only keys expose read tools only. Refresh preserves unsaved edits, and the page provides connection configuration and a built-in guide.
• Standalone runtime: Install and run the static Go binary without Node.js, npm or a Go compiler. Management UI, help, icons and the graphical Pipeline editor are embedded.
• Installation guides: Use package installation, startup, listener, upgrade and licensing instructions throughout the five in-product languages and the Chinese/English public guides.

Validation and upgrade

Both Linux architectures passed installation, login, embedded-resource checks, native Pipeline processing, real HTTP forwarding and restart recovery in isolated offline Alpine environments without Node.js or build tools. Upgrading retains existing configuration, queue data and credentials; back up and stop the service before installing.

0.1.3 · 2026-10-11

Reduce delivery overhead and synchronous acknowledgement costs while preserving durable, at-least-once delivery.

• Protobuf forwarding: Validate the SDK wire format and send the original bytes without constructing discarded Point objects. Unknown fields, numeric values and timestamp precision are preserved.
• Batch delivery: Use Redis pipelines, PostgreSQL batch INSERTs inside the existing transaction, AMQP confirmation windows of up to 64 messages and Pub/Sub SDK batching. Check all results and retain rollback, NACK and unroutable-message failure behavior.
• Connection reuse: Reuse clients across delivery rotations and short empty-queue periods, isolated by flow, destination and immutable version. Cache at most 32 clients and close them on errors, version changes and shutdown.
• Durable acknowledgements: Commit up to 64 concurrent acknowledgements in one synchronous journal transaction, including counters and ordered collection cursors. Return success only after commit. Encode and sync checkpoints outside the write lock; reclaim sparse segments in bounded steps while retaining newer transactions.
• Deployment guides: Update storage, delivery and operational documentation, plus listener addresses, public access, HTTPS proxies and container port mappings in all five in-product help languages.

Measured results

OrbStack Linux arm64, Go 1.26.7, 2 CPU. Medians from five before/after microbenchmark runs and six alternating end-to-end runs.

The end-to-end path was DataKit → Pipeline → combined routing → two Dataways, HTTP Protobuf and HTTP NDJSON. Each destination received 64,020 verified unique IDs per run, with matching digests, tags, large integers and nanosecond timestamps; duplicates, final backlog and failures were all zero.

Microbenchmark gains do not translate directly to whole-flow throughput. Tail latency and peak memory showed no clear improvement. These local measurements are not production capacity or long-run qualification.

Validation and upgrade

Source-package tests, vet and race checks passed, including real Redis, PostgreSQL and RabbitMQ batch/failure integration and local Pub/Sub SDK delivery. No configuration or queue-format migration is required. Synchronous durability, AES-GCM validation, immutable version snapshots and independent destination retry remain in place.

Back up data and credentials, stop the service, then install the new package. Detailed performance and validation records are included in the packaged documentation.